Quick post about a change I made that’s worked out well.
I was using OpenAI API for automations in n8n — email summaries, content drafts, that kind of thing. Was spending ~$40/month.
Switched everything to Ollama running locally. The migration was pretty straightforward since n8n just hits an HTTP endpoint. Changed the URL from api.openai.com to localhost:11434 and updated the request format.
For most tasks (summarization, classification, drafting) the local models are good enough. Complex reasoning is worse but I don’t need that for automation workflows.
Hardware: i7 with 16GB RAM, running Llama 3 8B. Plenty fast for async tasks.
I’m not a huge fan of AI, but I consider myself pretty open minded and have been considering doing a demo of Claude to at least gain an understanding of the tech I’m constantly talking shit about.
Is there anything self-hostable that compares in quality to what vibe coders claim Claude Opus is capable of?
The trash talking on AI is half people with legitimate concerns on the societal and ecological impact and the other half just want to be in on the party and aren’t interested in understanding it. It’s useful like googling things is useful, the items you search for are not always correct, but if you have a basic level of knowledge it’ll help you get where you want to be much faster.
Nothing quite compares to Claude Opus in a cohesive package that I’d recommend for an average self hoster but I personally really like running Nemotron from Nvidia. It’s not the best model, but in my experience it’s consistently good enough along with being fast and stable. If you’re focused more on coding, I hear the Qwen series had some good models.
I actually did an experiment on doing just that. For context, I’m an experienced software engineer, whose company buys him a tom of Claude usage so I had time to test out what it can actually do and I feel like I’m capable of judging where it’s good and where it falls short at.
How Claude Code works is that there are actually multiple models involved, one for doign the coding, one “reasoning” model to keep the chain of thought and the context going, and a bunch of small specialized ones for odd jobs around the thing.
The thing that doesn’t work yet is that the big reasoning model has to still be big, otherwise it will hallucinate frequently enough to break the workflow. If you could get one of the big models to run locally, you’d be there. However, with recent advances in quantization and MoE models, it’s actually getting nearer fast enough that I would expect it to be generally available in a year or two.
Today the best I could do was a tool that could take 150 gigs of RAM, 24 gigs of VRAM and AMD’s top of the line card to take 30 minutes what takes Claude Code 1-2. But surprisingly, the output of the model was not bad at all.